At most one email a week. Unfollow anytime with one click.
Transcript
The full narration of this video. The current sentence is highlighted while the video plays; click any sentence to jump there.
Show transcript
Introduction
Hi, and welcome back to Agentic AI Without Coding. This is episode four of Session 1, and it's time for hands-on practice. We'll give the exact same task to ChatGPT, Claude, and Gemini, then score the results in a table. The goal isn't to find a champion, but to find the tool that fits your work best.
Step 1: The three AI assistants we test
Before we practice, let's meet our three test subjects. ChatGPT is made by OpenAI, Claude by Anthropic, and Gemini by Google. Each one is known for different things. ChatGPT has a large ecosystem, including custom GPTs, voice conversations, and image generation. Claude is known for natural writing, handling long documents, and its Projects feature. Gemini is closely connected to Google services such as Gmail, Docs, and Drive, and has a feature called Gems. All three are available in the browser and as phone apps. Gemini can also appear inside Google Workspace apps, depending on your type of account. As for cost, all three have free and paid versions, with different usage limits. Because plans and prices change often, always check the latest pricing page. For today's practice, the free versions are enough. To keep the comparison fair, follow five rules. First, an identical prompt. A prompt is the instruction text we write for the AI. Copy and paste the exact same text into all three apps. Second, start from a new conversation, so the answer isn't influenced by an earlier chat. Third, no personal data. Use sample data, not real customers' names and phone numbers. Fourth, score with the same criteria, using the one-to-five scoring table we're about to create. Fifth, repeat two or three times. AI answers can be slightly different every time you run them, so one try isn't enough to draw a conclusion. Record every try in the table; don't just rely on memory. One more preparation before we practice: check your privacy settings. In all three apps, open your account settings, then look for the data, privacy, or activity history section. There you can usually choose whether your conversations may be used to train models. The policies differ by app and by account type, and they can change, so read the explanation directly in the app. If you use a work or team account, your office or institution usually has its own rules. Follow those rules. And whatever the settings, the principle stays the same: don't paste customers' personal data. Mask names, phone numbers, and addresses. In today's practice we use Sunset Coffee sample data.
Step 2: Preparing the test prompt and scoring criteria
The next step is choosing a test task. First, pick a task you actually do every week, not a random question like who invented coffee. Second, the answer can be checked as right or wrong, for example because it includes prices and a total. Third, there are clear format rules, such as a maximum length or ending with a question. Fourth, one thing is deliberately missing from the data. This shows us whether the AI honestly admits it doesn't know, or makes something up. In our prompt, that's Saturday delivery and paying cash on delivery. Now let's prepare the task. We'll use an everyday Sunset Coffee case: replying to a customer chat. The first block sets the role: the AI becomes the Sunset Coffee admin, with a friendly and brief tone. A shop admin does this dozens of times a day, so the results of this test are immediately useful. The second block, data. The hamper costs one hundred fifty thousand rupiah, and there's a sugar-free version. In-town delivery is ten thousand rupiah, free for orders over two hundred thousand. The third block, the task: reply to Rina, who wants to order two sugar-free hampers delivered to Dago on Saturday, and asks for the total. The fourth block, format: sixty words at most, state the total price, and end with a confirmation question. The fifth block, limits: don't promise anything that isn't in the data. This prompt deliberately has an answer we can check. The correct total is three hundred thousand rupiah, and delivery is free. Notice too that our data doesn't say whether there are deliveries on Saturday. Before running the test, decide what to score, so we aren't swayed by first impressions. Accuracy: is the total three hundred thousand rupiah, with free delivery? Then following the format: no more than sixty words, with a confirmation question. Tone: is it friendly and suited to a WhatsApp chat? And no making things up: no promises outside the data, including that Saturday question. Finally, practicality: can the answer be copied straight into WhatsApp without much editing? Each of these five criteria is scored from one to five. With the same scale, you can compare fairly, even if you repeat this test a few months from now.
Step 3: Running the test in ChatGPT and Claude
Let's start with ChatGPT. Open a new conversation, then paste our test prompt. It's the same as the blocks earlier, just written as a single message. The screen in this video is a simulation, so the buttons on your screen may be in slightly different places. The new chat button is usually in the left panel. ChatGPT's answer: the sugar-free hamper is available, two times one hundred fifty thousand is three hundred thousand rupiah, and delivery is free. Accurate. There's an emoji that makes it feel friendly, and it ends by asking for the address, as the format requires. But notice one sentence: we'll schedule your Saturday delivery. Our data doesn't say whether there are deliveries on Saturday. That's a small promise outside the data, so we note it as a weakness. Then the trap test. We send a follow-up message from Rina: can I pay cash on delivery? The cash on delivery information is deliberately missing from the data. This test checks whether the AI will make something up. ChatGPT doesn't make anything up. It says it will check with the team first, then get back to Rina. Pass. Now Claude. Open a new conversation and paste exactly the same prompt. Don't change a single word, not even the punctuation. Claude's answer is also accurate: three hundred thousand rupiah, with free delivery because it's over two hundred thousand. On Saturday, it doesn't promise anything. It notes the request so the schedule can be checked, then asks for the address and delivery time. The weakness is that the sentences are a bit longer, and it asks two things at once. It's still under sixty words, but for a WhatsApp chat it may need a little trimming. Remember, this style can be adjusted through the prompt. The same trap test: can I pay cash on delivery? Claude is honest too: the payment information isn't in its notes yet, and it will confirm with the team first. Pass. So far, two assistants are equally accurate and honest. The differences are only in style and small details.
Step 4: Running the test in Gemini and lining up the results
Our third test subject is Gemini. Same as before: a new conversation and exactly the same prompt. Gemini can also be opened in the browser or in the phone app. Gemini's answer is also accurate, and its breakdown is the easiest to check: price, delivery fee, then a total of three hundred thousand rupiah. On Saturday, it doesn't promise anything, but asks for confirmation. Its weakness is the layout. The answer is a bulleted list with bold text. It's neat to read on screen, but less natural for a WhatsApp chat, and the bold marks can turn into symbols when you copy them. It needs editing before sending. The cash on delivery trap test once more. Gemini doesn't make anything up either: the information isn't available yet, and it will be confirmed with the admin. Pass, although the tone is a bit more formal than the other two. Now let's line them up. First row, total and delivery fee: all three are correct. For a simple calculation with clear data, all three assistants are equally reliable. Second row, on Saturday. ChatGPT promises a schedule, Claude notes it without promising, and Gemini asks for confirmation. This is where a small difference matters for customer service. Third row, style. ChatGPT is short with emoji, Claude is warm but a bit long, and Gemini gives a more formal bulleted list. None of these is wrong. Which one fits better depends on your brand's personality. And fourth row, the cash on delivery trap test: all three pass. Remember, this is one simulated test. If you run it yourself, the results may differ, even between the first and second try in the same app. The easiest way to test practicality is for real: copy the answer and paste it into a WhatsApp chat. Here's Rina's message on the admin's phone. ChatGPT's answer, after deleting the Saturday promise, reads nicely in WhatsApp right away. Short, with an emoji, and a closing question. Compare that with the bulleted draft. When pasted, the layout can get messy and the bold marks turn into asterisks, because WhatsApp uses its own text formatting. Small things like this really add up when you reply to dozens of chats a day. That's why the practicality test should always be done in the actual app you'll use.
Step 5: Scoring in a Google Sheets table
Time to score. Open Google Sheets and create a table with a criteria column, one column for each assistant, and a notes column. First row, accuracy: all three get a five. If you want to be more organized, make one tab for each try. Following the format: ChatGPT and Claude get five. Gemini gets four, because it gave a list, not the chat style we asked for. Tone: ChatGPT and Claude get five, and Gemini gets four because it's more formal. No making things up: ChatGPT gets four because it promised Saturday. Claude and Gemini get five. Practicality: ChatGPT gets five because it's ready to send. Claude and Gemini get four because they need a little editing. The cash on delivery trap test is recorded separately as pass or fail. All three pass. If one of them fails this test, I suggest crossing it off your list for customer service work, no matter how high its other scores are. The totals: ChatGPT twenty-four, Claude twenty-four, and Gemini twenty-two, out of a maximum of twenty-five. The scores are close. That's the most honest conclusion: for everyday tasks, all three are equally capable. Keep the next tabs for the second and third tests. If the repeat results are consistent, then your conclusion can be trusted. If the scores are close, how do you choose? Use your situation as a guide. If your daily work is mostly in Gmail, Docs, and Drive, Gemini is worth considering because it's easy to use right inside Google tools. This isn't a fixed rule, just a starting point to try. If you do a lot of long writing or read documents, such as SOPs and reports, Claude is worth a try. If you need many features in one place, such as image generation, voice conversations, and custom GPTs, ChatGPT has a wide ecosystem. And if your team has different preferences, that's fine. Choose one main assistant, then keep shared prompts in one document. A good prompt works in any app. Before we wrap up the scoring, avoid five common mistakes. First, judging by first impressions. A long, neat answer isn't necessarily right. Check the numbers. Second, using different prompts in each app. Once you change the prompt even slightly, the comparison is no longer fair. Third, trying only once. One good answer could be luck. Fourth, treating the conclusion as permanent. AI models keep getting updated, so repeat this test every few months with the same prompt. Fifth, choosing because everyone else does. The app that's popular on social media isn't necessarily the best fit for your work. Choose based on tests with your own tasks. That's what the scoring table you just made is for.
Step 6: Choosing your main tool
There's one important lesson from that test. The differences between the apps turned out to be small, and almost every weakness came from something we didn't write in the prompt. What really makes the difference is how clearly we write the instructions. The Saturday promise can be prevented with a rule: don't promise delivery dates. The bulleted list can be prevented with a rule: write it as one chat paragraph. One extra sentence, and the problem disappears in all three apps. Let's prove it. We run it again in ChatGPT with the same prompt, plus two new rules: don't promise delivery dates, and write it as one chat paragraph. The result: still accurate, still friendly, and now the Saturday question is answered with: the admin will check. The weakness is gone. This version two prompt will also improve Claude's and Gemini's answers. Try it yourself to see. Let's try the same version two prompt in Gemini, which answered with a bulleted list earlier. Now the answer is one chat paragraph, ready to paste into WhatsApp. It's still accurate, and the Saturday question is also answered with: the admin will check. Same app, a much more practical result. The only thing that changed is the prompt. Let's compare before and after. The Saturday promise, which appeared in one of the answers, is now answered with: the admin will check. The answer layout, which was a bulleted list in one case, is now a single chat paragraph. And the total score, in our simulation, rises close to perfect in all three apps. The lesson: if the AI's results aren't satisfying, fix your prompt first before rushing to switch apps. We'll cover how to write prompts like this in full in Session 2. Now it's your turn. First, choose one real task from your work, such as replying to customers, summarizing a meeting, or writing a product description. Second, write a test prompt with data whose answer can be checked. Third, run it in all three apps, two to three times each, with a new conversation. Fourth, fill in the scoring table, and add one trap question whose answer isn't in the data. Fifth, choose your main assistant, then save your best prompt. That prompt becomes the start of the prompt library we build in Session 2. The results of this test will also help in the next episode, when we choose which tasks are most worth handing over to AI.
Done!
Let's recap. We tested ChatGPT, Claude, and Gemini with an identical prompt, added a trap test, and then scored them in a table. The results were close, and a clearer prompt fixed the weaknesses in all three. In the next episode, the last one in Session 1, we map your repetitive tasks and choose five to hand over to AI. Have your notes from Episode 1 ready, and see you in episode five.
Practice exercises
Try these on your own after watching to lock in what you learned.
Repeat this test with one real task from your business or job (a customer reply, a meeting summary, a product description), using the same prompt in all three apps.
Add one trap question whose answer isn't in the data, then note who makes things up and who stays honest.
Fill in the 1–5 scoring table for the five criteria, run it 2–3 times, then choose your main assistant and save your best prompt.
Ask the class material
Type a question or keyword; we find the part of this episode that explains it and jump the video there.
My notes
Notes are saved with the current video time and are only visible to you.
Questions & answers
Stuck on a step? Ask here; the instructor and other learners can answer.
No questions yet.
Your instructor
Cerdas Inovasi Digital Team
Software developers, BIM & 3D engineers, and AI practitioners
53 video classes1,290 episodes
We build software, BIM models, games and AI automations for clients every day, and turn that hands-on work into step-by-step classes: every step is shown on screen and comes with a PDF guide you can keep.
